New Requirements for Artificial Intelligence in GMP Manufacturing
With Annex 22, the EU GMP guidelines gain a dedicated chapter for AI and machine learning models in GxP-critical applications for the first time. It supplements Annex 11 (Computerized Systems) with specific requirements for AI and ML models used in GMP-critical processes, that is, in any scenario where an AI decision directly impacts patient safety, product quality, or data integrity. For pharmaceutical manufacturers, this means that existing and planned AI systems must now undergo a review.
What Annex 22 specifically requires
Annex 22 is currently still in draft form but is set to be finalized soon, meaning implementation will shortly become necessary. The draft outlines a risk-based lifecycle approach for AI models in GxP applications, covering everything from development and validation to ongoing operation.
An overview of the key components:
Intended Use & Risk analysis
Intended use, application limits and the consequences of failure must be documented and approved by subject matter experts before testing begins.
Validation with independent test data
Training, validation and test data must be strictly separated (four-eyes principle) to rule out bias and overfitting.
Explainability & Confidence Scores
Methods such as SHAP or LIME, as well as logged confidence values for each prediction, including an "undecided" designation in cases of low certainty.
Human Oversight
Qualified personnel must review and document AI-supported decisions in a transparent manner, scaled according to criticality.
Change Control & Drift-Monitoring
Every change to the model, system or process is evaluated; model performance is continuously monitored during operation.
Comprehensive documentation
From data characterization and test plans to deviations integrated into existing Annex 11 and GAMP 5 processes.
Taking stock instead of waiting and seeing
"Existing GxP-critical AI systems must be retrospectively reviewed and adapted to the new requirements following the entry into force of Annex 22."
Even though the final text of Annex 22 has not yet been adopted, the current draft does not provide for a transition period for AI systems already in production use. Anyone currently using an AI-based inspection system in GMP manufacturing should anticipate that a retrospective gap analysis and adjustment to the new requirements will be necessary once Annex 22 comes into effect.
Moreover, the current situation is such that existing GMP obligations (including Annex 11) already allow inspectors to scrutinize the use of AI systems. Early preparation prevents the pressure to make last-minute adjustments, production interruptions, and the need for retrospective documentation under time constraints.
Annex 22-compliant AI solutions from Hamburg – available today
As a long-standing partner to the pharmaceutical industry, Krempien+Petersen develops and validates AI-based image processing solutions that are already designed to meet the requirements of Annex 22, including comprehensive test data recording, static and validated models, and end-to-end traceability.
Glass breakage inspection
Optical inspection of vials and syringes using deep learning reliably detects cracks, fractures and chips in the glass body at speeds of up to 650 units per minute. It already meets the requirements of EU GMP Annex 22 and FDA 21 CFR Part 11.
Go to the glass breakage inspection product pageDeep Learning / KUPvision
Static, trained deep learning models for OCR, assembly inspection, surface defect detection and classification, integrated into KUPvision software with full documentation for validation and change control.
Go to the Deep Learning product pagePharmaceutical industry
Inspection and identification solutions for labeling, fill levels, blisters, OCR/OCV, code reading and glass inspection, tailored to the strict quality and safety requirements of the pharmaceutical industry.
Go to the pharmaceutical industry pageQualification & Validation
Support with test planning, test data independence, documentation and the qualification of your existing or new AI systems in preparation for Annex 22.
Go to the Documentation and Qualification page
Achieving Annex 22 Compliance in 5 steps
Identify all GxP-critical AI/ML systems in use, including legacy systems and third-party software.
Comparison of the existing validation with the requirements of the Annex 22 draft (intended use, test data, explainability, oversight).
Retrofit missing evidence (e.g., test data independence, confidence values, risk analyses).
Define responsibilities, training and testing protocols for AI-supported decisions.
Implement change control and ongoing drift monitoring to continuously demonstrate model performance.
